EXANTE’s Codey hits 75% success, but humans keep control

EXANTE’s Codey hits 75% success, but humans keep control

Investment firm EXANTE has spent six months embedding an autonomous AI developer into its engineering teams. Its experience suggests that the real barrier to AI-led software delivery is not model capability but organisational discipline.

According to EXANTE, its head of web applications, Alina, led the initiative as part of a wider AI transformation programme. Her team built Codey, an in-house multi-agent system that takes a task from a Jira ticket through to a working feature with little human intervention.

Since launch, 250 tasks have passed through the system and 75% were completed successfully. Codey now handles 7% of all work in the teams that use it, and typically moves from the start of coding to a deployed feature environment in around 30 minutes.

EXANTE’s journey began with narrower automation, such as AI code review and analysis of the modules affected by code changes. These tools sped up individual stages but stopped short of full-cycle development. Codey was designed to close that gap. It runs on Claude Agent SDK, Codex and OpenCode, and coordinates six specialised agents: a researcher, planner, developer, reviewer, QA tester and verifier. Different models are assigned to different roles, and work flagged by the reviewer or automated checks is sent back for revision.

The firm now uses Codey across projects ranging from internal CRM and back-office systems to client-facing FinTech services, including its client portal, web trading terminal and desktop application. Tasks can arrive through sprint planning, where teams share a backlog with the system, or through Slack, where Codey clarifies requirements, raises a Jira ticket and returns with a merge request and preview link.

Safeguards are built in throughout. Codey is limited to three revision cycles before a task is marked as “AI failed”, and humans retain responsibility for product acceptance and release decisions. EXANTE says most failures stem from missing context, changes spanning multiple interdependent repositories, or first tasks in an unfamiliar project or stack.

Perhaps the most striking finding concerns model selection. EXANTE’s benchmarks showed that using the most powerful model at every stage does not always deliver the best outcome. Simpler models sometimes follow plans more faithfully, overcomplicate less and finish faster, prompting the firm to move some roles to Codex models. A “Learning loop” tests improvement hypotheses against fixed benchmarks; in one cycle, nine of 20 were kept.

Challenges remain, including the need for well-written requirements, building team trust, a temporary shift of workload towards review and the infrastructure cost of feature environments.

Looking ahead, EXANTE plans automated backlog analysis, wider departmental coverage, richer context sources and feature-based delivery, with the ultimate aim of handing Codey the full cycle from task definition to production.

Read the daily FinTech news

Copyright © 2026 FinTech Global

Enjoying the stories?

Subscribe to our daily FinTech newsletter and get the latest industry news & research

Investors

The following investor(s) were tagged in this article.